Evaluating the Impact of Local Data Imbalance on Federated Learning Performance for IoT Anomaly Detection

The Internet of Things (IoT) has experienced tremendous surge in the number of connected devices, resulting in an enormous amount of data being generated. Anomaly detection in such scenarios is crucial for ensuring the security and reliability of IoT systems. Federated Learning (FL) has been proposed as a viable solution for anomaly detection in IoT, allowing distributed devices to collaboratively learn a global model while keeping their data private. In this paper, we investigate the impact of local data imbalance on the performance of FL-based anomaly detection models in IoT scenarios. In general, data imbalance in FL can manifest either globally, where the collective data of clients is imbalanced, or locally, where the data of individual clients is imbalanced. To specifically examine the impact of the later, we use the X-IIoTID dataset and create three different data distributions, each consisting of 10 client devices. The global data imbalance is kept similar across all distributions but the level of local data imbalance is varied. We then apply FL using a federated averaging algorithm to train an anomaly detection model on each distribution. Our study reveals that despite having the same global data imbalance, an increase in the level of local data imbalance results in a decline in the FL-based model’s performance especially when the client devices train on different sets of data classes. Our findings suggest that addressing local data imbalance is crucial for achieving accurate and robust anomaly detection in IoT using FL.